The value-add of deep learning in predictive maintenance

As we enter the 4th Industrial Revolution – what the World Economic Forum describes as “bringing together digital, physical and biological systems” – the machinery that drove the third is beginning to integrate AI technology in order to render manufacturing and maintenance processes more efficient.

Every machine degrades over time. Predictive maintenance (PdM) should be a core process to avoid downtime. Deep learning is a great potential technique to apply to PdM because it’s good at identifying patterns in scenarios involving large, complex datasets containing multiple types of data. In addition, by adding audio or image data to tabular/sensor data, PdM can improve the models on machine or service degradation. In a time when manufacturing companies are under intense pressure to improve efficiency and productivity, every gain is valuable.

Unplanned maintenance can have a huge financial impact on organizations when you tally up the costs
— Rob Dalgety, Product Marketing at Peltarion

“In our increasingly digitized world, where virtually every activity creates a digital trace, there has been exponential growth in how much data can be used for predictive maintenance.” – PwC

The value-add of predicting maintenance

Most manufacturing companies deal with two basic types of maintenance: planned and unplanned. Unplanned maintenance can have a huge financial impact on organizations when you tally up the costs of the maintenance itself in addition to factory-line disruption and having a fixed asset unutilized for a period of time.

Obviously, the goal of any company is to reduce the number of incidents of unplanned maintenance by carefully controlling the occurrence of planned maintenance. The types of planned maintenance most companies use can be ordered by increasing sophistication:

1. Preventative maintenance
2. Corrective maintenance
3. Predictive maintenance

How deep learning adds even more value to manufacturing operations

“There’s one significant asset that manufacturers have not yet optimized: their own data. Process industries generate enormous volumes of data, but many have failed to make use of this mountain of potential intelligence.” – McKinsey

Most modern factories are already outfitted with machinery sensors. Deep learning can add power to predictive maintenance data in two ways:

1. By adding multiple types of data to the models: images or audio or video, on top of existing sensor data, for an enhanced dataset that powers a comprehensive predictive model.
2. Unlike other methods, typically, the more data you can provide, the more deep learning models improve, and the more accurate the model becomes.

For example, audio data, in particular, is a powerful source of data for predictive maintenance models. Sensors can pick up sound and vibration and used in the deep learning machine learning models. This data can be the most critical in determining machine life and degradation cycles.

Deep learning applied to predictive maintenance could have practical applications for many types of assets, for instance:

Reducing operating costs, and increasing reliability and productivity on factory floors

Monitoring complex power grids that contain enormous numbers of diverse assets such as transformers, cables, turbines, storage units and more

Identifying the risk of equipment failure in oil and gas pipelines and drilling equipment

Predicting degradation to pipes, power lines or cabling

Bolstering aging infrastructure within water utility systems

Assessing performance in large solar farms with million of modules and hundreds of millions of cells

The path to using advanced AI techniques is more realistic and realizable than many businesses think with operational AI
— Rob Dalgety, Product Marketing at Peltarion

For example, in the railway industry, deep learning is already enhancing predictive maintenance for companies like the Belgium-based Infrabel to predict mechanical failures in switches before they occur. By putting in place sensors, cameras and meters, Infrabel is more efficiently and economically able to predict overheating and drifts in power consumption.

And for manufacturers of metal and plastic products, such as Mueller Industries in Tennessee, USA, predictive maintenance boosted by deep learning aids in conducting sound analysis on machines to predict impending malfunctions.

The current state of deep learning for predictive maintenance

Half of the respondents to a recent PwC survey about predictive maintenance using machine learning said they plan to implement such systems within the next five years – although only 11% have currently achieved this level. Within these machine learning numbers, that likely implies that a very small percentage are using deep learning.

Given the power of deep learning and the rapid advancement in the solutions which underpin deep learning, this is a missed opportunity. Deep learning techniques are available to organizations of all types today. The path to using advanced AI techniques is more realistic and realizable than many businesses think with operational AI.

Is your organization is ready to embrace AI and deep learning? Interested in experimenting with deep predictive maintenance for deep learning applications in a hands-on way?

Check out our upcoming webinars, or recordings of past events, here.

Rob Dalgety
Product Marketing Director

About

Rob Dalgety leads Product Marketing at Peltarion. He has extensive experience in commercializing and positioning software into enterprises and other organizations in areas including mobility, big data and analytics, collaboration and digital.

More from the author

Get started for free